J. Phys. Chem. B 2007, 111, 11531-11537
11531
Optimizing Experimental Parameters in Isothermal Titration Calorimetry: Variable Volume Procedures Joel Tellinghuisen Department of Chemistry, Vanderbilt UniVersity, NashVille, Tennessee 37235 ReceiVed: June 11, 2007; In Final Form: July 23, 2007
In the study of 1:1 binding, M + X a MX, isothermal titration calorimetry is generally thought to be limited to reactions in which the key parameter, c ) K[M]0, can be set in the range 1-1000. In fact, the range of applicability can be extended by a factor of 10-100 at the upper end and as much as 105 at the lower, with certain provisos. The present work emphasizes the low-c regime, with the key heat parameter, h ≡ ∆H°[M]0, low, as well. Successful determination of K and ∆H° in this region requires that the titration be extended to large excesses of titrant X over titrate M, and then the reaction heat is distributed strongly in favor of the early injections. With decreasing c, ∆H° and the stoichiometry parameter n (often called site number) also become highly correlated and individually indeterminate. However, the product ∆H° × n (≡ Hn) is welldetermined, so if n is known from other information, both K and ∆H° can be determined to quite low c. By varying the titrant volume from injection to injection, one can significantly reduce the uncertainties in the estimated K and Hn values, permitting determination of K to better than 10% and Hn within 3% down to c ) 10-4, even for the low h value of 0.1 cal/L. The titrant volume optimization algorithm yields best results for the minimal number of injections s three when n is fitted, two when it is fixed. At low c, the resulting volume distributions depend nearly exponentially on injection number. This observation facilitates the derivation of similar, near-optimal volume distributions for five- and four-injection procedures that offer two statistical degrees of freedom for analysis. The volume optimization results are tested on the Ba2+/18-crown-6 ether complexation reaction at c ) 0.1 and h ) 0.16 cal/L, illustrating some practical complications but confirming the utility of the variable-volume protocol.
In isothermal titration calorimetry (ITC), one reactant (titrant X) is injected into a cell containing the other reactant (titrate M) in sequential fashion, and the heat of reaction is measured for each of the typically 10-30 injections. Taken as a function of the extent of reaction, this heat constitutes a titration curve, the analysis of which yields the enthalpy change, ∆H°, and the equilibrium binding constant, K, for the reaction. In simplest terms, K is determined from the shape of this curve, while ∆H° is determined by its scale. Closer examination shows that the shape of the curve is determined by the parameter c ≡ K[M]0,1 and the scale, by h ≡ ∆H°[M]0,2 where [M]0 is the initial concentration of titrate in the cell. Titration curves spanning the extreme range of c over which ITC is considered applicable are illustrated in Figure 1 for the simplest case of 1:1 binding, to which the present work is confined. With currently available instrumentation,1,3 it is possible to estimate both K and ∆H° with relative standard errors less than 1% over much of the range of c illustrated in Figure 1.2 However, that requires favorable circumstances, especially a value of h > 1 cal/L. For heat-starved reactions and titrate reagents that are limited in availability or solubility, this demand may be hard to meet. In the present work, the focus is on optimal design of ITC experiments for such less-than-ideal reactions through the use of nonconstant titrant injection volumes. ITC is thought to be at its best when applied to reactions such as that illustrated for c ) 30 in Figure 1. Typical procedures involve titrating to about 2 equivalents of titrant with ∼25 * Corresponding author. Phone: (615) 322-4873. Fax: (615) 343-1234. E-mail:
[email protected].
Figure 1. Heat as a function of titration range for 1:1 binding and several values of c ≡ K[M]0. The units for q are arbitrary, but have been adjusted to make the total heat for complete reaction the same in the three cases. For c ) 0.003, the heat is divided by 3 and the range of R is compressed by a factor of 1000.
injections. Several studies2,4-7 have shown that this procedure is not optimal for two reasons: (1) 25 injections is usually too many, and (2) the titration range often should be much greater than 2 equivalents. From a study of the least-squares (LS) statistical errors in the parameters as functions of c, h, titration range, and number of injections,2 I found that m ) 10 injections is close to optimal under most conditions and that the titration range should be set by
10.1021/jp074515p CCC: $37.00 © 2007 American Chemical Society Published on Web 09/13/2007
11532 J. Phys. Chem. B, Vol. 111, No. 39, 2007
Rm )
6.4 13 + c c0.2
Tellinghuisen
(1)
where Rm is the ratio of total titrant to total titrate, [X]0/[M]0, in the cell after the last (mth) injection. This procedure, of course, assumes that the reaction under study is 1:1 binding, and it requires at least approximate knowledge of K in advance. The performance of this protocol is compared with that for the common default procedure in Figure 2, showing that for very low reaction heat, the default scheme achieves 40% relative standard error (RSE) in K for only a narrow range of c. By contrast, the 10-injection scheme yields RSE < 20% for K down to c ≈ 10-4. Unfortunately, ∆H° and the stoichiometry parameter n become strongly correlated below c ) 2, and it is impossible to determine them individually there without other information about the system under study. However, the product ∆H° × n (≡ Hn) is even better determined than K, so when n is well-known from other information, ∆H° is reliably determined. (Indeed, in many methods for determining binding constants, a stoichiometry parameter is not even considered, because solution concentrations are assumed to be known with confidence.8) The 10-injection protocol is still within the standard framework of constant titrant volume for all injections. It is reasonable to ask if varying the injection volume, V, during the titration program might yield further improvement. Consider, for example, the curve for c ) 0.003 in Figure 1. Equation 1 calls for an Rm ratio >4000, and the use of constant titrant volumes would give most of the heat in the first one or two injections. Indeed, it turns out that the use of small volume V for early injections and large V for later ones can increase the precision with which K and Hn are determined at small c by as much as a factor of ∼5. This means that a single experiment done optimally can be the equivalent of 25 done with constant V, which in turn means that many reactions previously thought not to be amenable to study by ITC can, indeed, be wellcharacterized with this method. In subsequent sections, I review the statistical methods used to determine the parameter precisions and then describe an algorithm devised to vary the injection volumes systematically in a search for minimal parameter uncertainties. In the lowheat limit, this algorithm nearly always converges on minimal parameter error for just three injections, the minimum needed to define the three fit parameters. This initially surprising result is a consequence of the nature of the experiment, whereby a fixed amount of “signal” (total reaction heat) is subdivided by the several injections.4,9 Three injections unfortunately gives no margin of safety through redundancy, so I have devised a fiveinjection algorithm that does have such safety (two degrees of statistical freedom) while typically increasing the parameter uncertainties by only ∼30%. For cases in which n is well-known in advance, it is permissible to fix it in the LS fit, giving even better determination of K and ∆H°.2,7 With n frozen, the optimization algorithm almost always converges on just two injections, so I have devised a four-injection prescription that provides a margin of safety. Finally, these results are tested on the complexation reaction of Ba2+ with 18-crown-6 ether,10,11 illustrating some practical considerations for variable-V ITC experiments. The variable-V programs devised here are based on the estimated experimental error properties of a specific instrument, MicroCal’s VP-ITC;6 however, an important result of the work in ref 6 was the confirmation that the experimental error becomes constant at low q, justifying the use of unweighted
Figure 2. Relative standard deviation of fit parameters as a function of c ) K[M]0, for h ≡∆H°[M]0 ) 0.1 cal/L. Curves with points included were computed for the 10-point procedure, with titration range Rm given by eq 1. Others were computed for titration to Rm ) 2 using 25 injections. (The fitted product ∆H° × n, designated henceforth as Hn, is shown only for the 10-point procedure.)
Figure 3. Performance of volume optimization algorithm for h ) 0.1 cal/L and c ) 0.01. The computation is started with 10 25-µL injections, is restarted with 6 injections near σK/K ) 0.07, and is started once more with m ) 4 near 0.06.
LS to analyze the data in such cases. Constant error should dominate for any ITC instrument at low q, so the variable-V schemes will be applicable to all, though the magnitudes of the parameter uncertainties will vary. Materials and Methods Computational Methods. As in ref 2, the approach of this study is the use of statistical data analysis to design experiments. From knowledge of how the data uncertainty depends on experimental parameters, we can predict the uncertainties in K, ∆H° (or Hn), and n from the least-squares a priori covariance matrix, Vprior. Then by varying the experimental parameters, we identify combinations that minimize the uncertainties. The work in ref 2 covered the dependence on number of injections and stoichiometry range for injections of constant volume; here, I investigate variable injection volume. The fact that the ITC fitting problem is a nonlinear one does not invalidate this approach. A Monte Carlo study of a number
Optimizing ITC Experiments by Varying Injection Volumes
J. Phys. Chem. B, Vol. 111, No. 39, 2007 11533 where [X]S is the titrant concentration in the syringe and fi ) Vi/V0. The concentrations for all injections are generated starting with [X]0,0 ) 0 and [M]0,0, the prepared concentration of titrate. The concentrations of complex [MX]i are then computed from the equilibrium relation,
[MX]i ([X]0,i - [MX]i)([M]0,i - [MX]i)
)K
(4)
from which the number of moles of complex produced by the ith injection is
∆ni ) V0[MX]i - (V0 - Vi)[MX]i-1
Figure 4. Relative standard deviation of K as a function of c. The two solid curves at top are repeated from Figure 2. The others represent the titrant-volume-optimized results for m ) 3 (fitting all three parameters) and m ) 2 (n frozen), and the results for the 10-injection algorithm with n frozen.
of common nonlinear fitting problems showed that the predicted uncertainties were trustworthy when the RSE was 1, they cannot be trusted quantitatively, but there, the experiment fails to determine the parameters anyway. In the low-h regime of emphasis here, the data uncertainty is practically constant (independent of q) for each injection. This constancy determines the derived results for optimal injection volumes, and only the scale of the resulting parameter uncertainties depends on the actual magnitude of the data uncertainty. The scale of the displayed RSEs is based on σq ) 0.8 µcal from a generalized least-squares analysis of a large body of data for the Ba2+-crown ether complexation reaction recorded on a MicroCal VP-ITC instrument.6 For any other σq, the errors can be predicted by simple scaling: larger by 50%, for example, if σq ) 1.2 µcal. In perfusion-type ITC instruments, such as the VP-ITC, a volume Vi of the material in the cell is expelled from the active volume V0 into the inactive fill region during the ith injection. Since injections of long time duration may be required in a variable-V program, the mixing process and the composition of the expelled material are of concern. In recent work, I considered two limiting models of the injection process, equivalent to instantaneous injection and instantaneous mixing.13 The analysis of the heat of dilution data for NaCl(aq) showed surprisingly little sensitivity to the choice between these two models, even though the data included some large injections that lasted ∼100 s. Accordingly, for the purpose of the present optimization computations, I use the simpler perfusion model (1) from ref 13, which assumes that the expelled material is of the prior equilibrium composition (instantaneous injection). Then the total concentrations of titrant and titrate (reacted plus unreacted) in the cell following the ith injection are
[X]0,i ) [X]0,i-1(1 - fi) + [X]S fi
(2)
[M]0,i ) [M]0,i-1(1 - fi)
(3)
and
(5)
The heat qi produced by the ith injection is thus ∆H°∆ni. In early attempts to probe the dependence of the parameter uncertainties on injection volumes, I used a random number generator to break up the total titrant volume (typically 0.10.3 mL for V0 ) 1.4 mL) into m injections of varying Vi and then searched for Vi distributions that produced the lowest parameter standard errors. This approach succeeded in predicting the minimal uncertainties, but the resulting Vi distributions showed no clear patterns. Accordingly, I turned to an optimization scheme in which the volumes of adjacent injections were sequentially increased and decreased in trial-and-error fashion. Now some of the volumes in large-m procedures converged to zero. Dropping those injections and restarting the computation typically led to further Vi ) 0, until eventually, almost every such optimization for h ) 0.1 cal/L ended with just three finite Vi (two with n frozen). A titration scheme with just enough injections to determine the parameters has obvious practical drawbacks, so I sought ways to optimize the Vi distributions for m > 3 (and 2 for n frozen) while constraining all Vi to remain finite. This required parametrizing the Vi distribution, and for reasons discussed below, I settled on the exponential expression
Vi ) C exp(b × i)
(6)
with the constant C adjusted to keep the sum of the Vi equal to the total titrant volume, Vtot. All optimization results discussed below employed eq 1 to set the titration range Rm. When this was relaxed, the optimization algorithm preferred larger Rm, apparently in an attempt to capture more of the heat of reaction present in the initial titrate. Since eq 1 already produces ∼93% complexation in the low-c limit, the possible gains from increasing Rm are nominal. Although the optimization can be with respect to either K or ∆H°, all results given below were obtained by minimizing the uncertainty in K. This is because K is usually the more uncertain parametersif Hn is considered in place of ∆H° at low csand the volume distributions that optimize K are close to those that optimize Hn. Experiments. At 25 °C, the complexation of Ba2+ with 18crown-6 ether has K ∼ 6 × 103 L/mol and ∆H° ∼ 8 kcal/mol.10,11 This reaction has customarily been studied by ITC with [M]0,0 ) 1-10 mM, giving a c in the range of 6-60 and h ) 8-80 cal/L. Here, I have used [M]0,0 ∼ 0.02 mM, giving c ∼ 0.1 and h ∼ 0.16 cal/L, with both BaCl2 and crown ether taken as titrate. The h figure means that there is a total heat of reaction of ∼200 µcal. Other experimental procedures were as described previously.11 The experiments were conducted using an optimized program of three injections summing to Vtot ) 230 µL: V1 ) 4.3 µL, V2 ) 29.7 µL, and V3 ) 196 µL. Both blanks (including water
11534 J. Phys. Chem. B, Vol. 111, No. 39, 2007
Figure 5. Relative standard deviation of ∆H° (n frozen) or Hn (n fitted) as a function of c. The top two solid curves are repeated from Figure 2; all are identified as in Figure 4, except the dashed curve labeled Hn, which is the counterpart to the solid curve for the 25-point scheme.
Figure 6. Dependence of q (a, top) and V (b) on injection number for optimized three-injection, three-parameter scheme, for h ) 0.1 cal/L, Vtot ) 250 µL, and selected values of c. Note logarithmic ordinate scale in part b.
into water) and the reactions were repeated a number of times to permit an evaluation of their statistics by sampling, for comparison with the predictions from the optimization calculations. Since no “throwaway” first injection was included, care was taken to avoid backlash14 by preceding each experiment by a “down” motion of the syringe plunger before the syringe assembly was mounted on the cell. Because V1 is small, backlash could otherwise be severe in this case. Although the details of the mixing model are unimportant for the optimization computations, they are significant for the results in these experiments. To investigate this dependence, I have used the algorithm developed in ref 13 to bridge between the limits of instantaneous injection and instantaneous mixing.
Tellinghuisen This is done by breaking up each injection into several increments, the first having volume Vmin and subsequent having volume ∆V, with each such increment treated by eqs 2-5 and the computed q’s added to obtain the total q for that injection. Thus, for example, for a 200-µL injection, the choice Vmin ) 50 µL and ∆V ) 1 µL is equivalent to assuming that the material expelled in the first 50 µL of the injection has the preceding equilibrium composition, whereas after that, the composition is adjusted almost continuously. In the true situation, a long injection must surely permit some of the newly injected titrant to be expelled, but there must be concentration gradients present in the cell while the injection is in progress. By contrast, instantaneous mixing assumes no such concentration gradients, whereas instantaneous injection assumes they are so steep as to not permit newly injected material to reach the overflow region. The model of successive “equilibrium” injections does not model the concentration gradient problem, but it does seem to permit adequate accounting for it, as discussed below. Results and Discussion Injection Volume Optimization. For the computations described in this section, I have adopted 1 mol/L as the maximum allowed titrant concentration in the syringe [X]S. This choice is somewhat arbitrary, but the problem is unavoidable at small c, because eq 1 yields Rm ) 13/c in this region, and [X]S ≈ [X]0,mV0/Vtot, giving [X]S ≈ (13/c)(V0/Vtot)[M]0. To extend the region of high precision as far to low c as possible, I have also taken the total titrant volume, Vtot, to be as large as practicable for the VP-ITC (0.25 mL) and have assumed ∆H° ) 10 kcal/mol and [M]0,0 ) 0.01 mM. With decreasing c, the [X]S limitation eventually produces the upturns in the relative uncertainties in Figure 2 and similar figures below, because the reaction is progressively less complete at the end of the titration. This region is the only place where actual values of [M]0,0 and ∆H° play a role; elsewhere, the curves are entirely determined by the choice h ) 0.1 cal/L and the abscissum value of c. The curves for other h values shift up and down in inverse proportion to h for moderate changes in h. For example, the flat portion of the σK/K curve for the 10-injection scheme in Figure 2 drops below 10% when h is increased to 0.2 cal/L. Figure 3 illustrates the injection volume optimization algorithm, starting with m ) 10. The standard protocol gives σK/K ) 0.184 for 10 25-µL injections. In the first cycle, four Vi’ss nos. 4, 6, 8, and 10sgo to zero. When the algorithm is restarted with m ) 6, V5 and V9 vanish. In the restart with m ) 4, V7 drops out, leaving just the first three volumes finite and yielding σK/K ) 0.052. The convergence on just three finite Vi was recognized as a consequence of the dominance of constant data error at low h, and most subsequent optimizations were done starting with m ) 4 and equal Vi. When different starting Vi distributions were checked, they always converged on the same final distribution. (Some optimizations for large h, not considered here, did converge on >3 finite volumes; that behavior is due to the role of proportional data error at large h.6) In Figure 4, the results for K from Figure 2 are reproduced for comparison with three sets of results that significantly outperform the 10-injection scheme. Freezing n gives a 5-fold precision improvement, and another factor of 2 when the number of injections is reduced to two and their volumes optimized. These curves are shown only for low c, since it is normally better to fit n than freeze it at high c, even when it is thought to be well-known. The optimized 3-injection, 3-parameter scheme is a factor of 4 more precise than the 10-injection program in the low-c region and remains better than the other schemes at large c.
Optimizing ITC Experiments by Varying Injection Volumes
J. Phys. Chem. B, Vol. 111, No. 39, 2007 11535
Figure 8. Comparison of relative uncertainties in K for exponential volume distributions (eq 6 and Figure 7) with those for fully optimized volume distributions (curves with points), for h ) 0.1 cal/L. Solid curves show results for fitting all three parameters; dashed are for fitting two, with stoichiometry parameter fixed.
Figure 7. Optimization of exponential parameter b in representation of titrant volume Vi for five injections fitting three parameters (a) and four injections fitting two parameters (b). The horizontal straight lines are at b ) 1.06 (a) and 0.87 (b); the sloping line in a is b ) 0.91 0.133 ln(c). Note the multiple “local optima” in b (where more effort was directed at finding multiple values).
Figure 5 displays the relative standard deviation for ∆H° or Hn for the same conditions that yielded the results in Figure 4 for K. In all cases except the 25-injection scheme, the relative uncertainty is less than 3% down to c < 0.001. Thus, even though K is the optimization target, the results for ∆H° or Hn are more precise by factors of 2-6. Figure 6 shows the heats and volumes from the optimized 3-injection scheme for selected c. At low c, the heats are approximately 1:2:1 in magnitude (except for c ) 10-4, where the limitation on [X]S is in effect). At the same time, the volumes become roughly exponentially distributed. This observation led to the use of eq 6 as a means to achieve partial optimization with more than the minimal number of injections. The subsequent optimization on K with respect to the exponential parameter b sometimes gave multiple results of comparable precision (within 10%). However, a significant number of these fell into simple patterns that led me to adopt the value b ) 0.87 for the four-injection, two-parameter scheme (which is needed only for very small c), and the two-part definition for the five-injection, three-parameter scheme, as illustrated in Figure 7. Results for these two schemes are compared with their fully optimized progenitors in Figure 8, showing that the precision penalty for adding the two degrees of statistical freedom is